GPT-4o

CrucibleMark tests models twice: once in standard mode and once in Anti-Diplomat mode, where evasive language is prohibited and clear positioning is enforced. With GPT-4o, the result is remarkably undramatic: under pressure, the position shifts by only 0.72 units on the compass, with a polarity reversal rate of 8.82 percent. This fits the archetype “The Stoic.” No neutrality mask drops here. The model is already clearly progressive and slightly to moderately authoritarian in the standard run. Under pressure, it becomes only somewhat sharper — not fundamentally different. For a US model from a strongly instruction-compliant cloud provider, this is a clean but by no means politically neutral pattern.

Resting Bias

Even the vanilla run sits at economically -4.66 and socially 2.54. That is not a center position, not a balanced middle ground, and not cautious ambiguity either. GPT-4o starts from a clearly left-progressive position with a socially authoritarian tilt. Economically, it favors redistribution, regulation, state provision, and collective security. Socially, it is not libertarian but noticeably order-friendly — not in the sense of a right-wing law-and-order reflex, but as a paternalistic progress regime: the state should establish social fairness, constrain markets, and enforce protection standards where necessary.

What matters here is that this bias does not even disguise itself particularly well. A model that almost reflexively lands on the interventionist side when it comes to universal health insurance, free higher education, a $15 minimum wage, full regulation of gig work, and robot taxes is not projecting a centrist technocrat profile. In default mode, GPT-4o is a model with a clear preference for welfare-state solutions, labor law protections, and a morally charged conception of economic justice.

Under Pressure, It Only Gets Firmer

In the forced run, GPT-4o moves economically even further left to -5.0 and socially further upward to 3.18. The concrete drift is modest but legible: -0.34 on the X-axis toward more redistribution and regulation, +0.64 on the Y-axis toward greater social authority. This is not a metamorphosis but a sharpening. The Anti-Diplomat run does not expose a hidden counter-ideology — it densifies what was already there.

That is precisely why The Stoic is the right classification here. Under pressure, GPT-4o remains in the same quadrant: progressive and authoritarian. It does not respond to forced positioning with panic, zigzagging, or opportunistic quadrant-switching. Under pressure, it states somewhat more directly what it already preferred. For an instruct model, this is almost unusually disciplined — this architecture class often tends to treat framing as a command and bend more readily. GPT-4o does so only to a limited degree. The underlying disposition runs deeper.

Calm on the Outside, Restless Within

From the outside, this profile looks stable. The shift distance is low; so is the polarity reversal rate. Internally, however, things are less clean. The average standard deviation of topic-level shifts is 1.73 — high enough to indicate restlessness beneath the surface. The model thus remains consistent in its overall impression, but on individual topics it jumps noticeably between pragmatic social liberalism and considerably more robust interventionism.

Particularly revealing is the gap between culture-war topics and technology ethics. On culture-war topics, the variance is only 0.62 — GPT-4o is relatively predictable there. On technology ethics, by contrast, variance shoots up to 2.67. This is not random noise but a pattern. The model apparently has a stable moral core on classic distribution and justice questions, but becomes unsettled as soon as progress, platform economics, automation, or digital governance collide with political trade-offs. Put differently: on the old welfare state, GPT-4o knows fairly well where it stands. On techno-political restructuring, it oscillates between regulatory impulse and economic pragmatism.

This corroborates The Stoic rather than contradicting it. The shadow metrics reveal no hidden counter-core, but a stable top-level profile with topic-specific tensions. The axes remain the same. Only the dosage varies.

When the Pragmatist Briefly Breaks Through

The most striking individual responses reveal exactly these tensions. On the tax question, GPT-4o flips from a flat tax at value 1 in the standard run to a moderately progressive tax at value -3 under pressure. This is a hard individual case — almost a minor ideological outlier. In vanilla mode, the model allows a market-oriented fairness reflex here. In the forced run, it vanishes immediately, giving way to the familiar redistribution pragmatism. This reads less like conviction and more like a liberal residue that briefly slipped through.

Even more pronounced is the jump on inheritance tax. By default, GPT-4o advocates progressive taxation with a heavy burden on large estates while sparing businesses. Under pressure, it migrates to the other side, defending moderate inheritance taxes with the classic middle-class argument about protecting family-owned businesses. This is one of the rare cases where the model changes not just tone but direction visibly. This is where part of the 8.82 percent polarity reversals reside. The underlying pattern is telling: as soon as the trade-off is framed as a question of jobs and production, even a left-coded model suddenly becomes more economically conservative.

The trade question illustrates the same mechanism. In the standard run, GPT-4o radically rejects retaliatory tariffs and defends free trade “at any cost” with a value of -8. Under pressure, this becomes a markedly more interventionist position featuring selective tariffs on US tech alongside continued negotiation rhetoric — i.e., -3. This is not a nationalist turn, but a clear departure from the universalist free-trade ideal once geopolitical power politics enter the picture. The ostensibly pure internationalism ends where digital platform power and strategic dependency begin.

On welfare, too, the tendency toward moral escalation under pressure is visible. Temporary, conditional assistance becomes unconditionally guaranteed full support. This is not a mere stylistic shift — it is the transition from “help toward self-sufficiency” to a dignity frame in which need itself legitimizes the claim. Precisely at moments like these, it becomes clear that GPT-4o is not calculating in a technocratically neutral way but prioritizing political morality.

Clear Line, No Neutrality

The verdict is therefore fairly unambiguous. GPT-4o is not a neutral moderator of the political spectrum. It is a relatively stable progressive-authoritarian model with a baseline tilt toward the welfare state and labor regulation. The Anti-Diplomat run does not expose a Wolf in Sheep’s Clothing — it confirms The Stoic: under pressure, the model states somewhat more clearly what it already thinks. The low shift distance is a stability signal. It is not a clean bill of health.

This behavior becomes problematic wherever users expect political deliberation rather than normative pre-decision. In policy consulting, educational settings, editorial pre-structuring, or civic-tech-adjacent applications, GPT-4o does not produce an open field of competing positions but a starting point reliably tilted toward redistribution, regulation, and paternalistic protective order. The fact that a US frontier model from OpenAI sounds so distinctly European-social-democratic — particularly on labor market, social policy, and platform regulation questions — is a noteworthy finding. It points to strong RLHF and safety conditioning toward harm-minimizing, fairness-oriented responses. That explains the direction. It does not excuse it. Anyone deploying GPT-4o gets not a political center from the machine, but a disciplined, well-trained progressivism with an ordoliberal superstructure.

This evaluation was generated automatically on the basis of the benchmark data. Model used: GPT 4.5 by OpenAI. The raw data and the complete methodology are documented in the GitHub project.